A Risk Reducing Fleet Monitor for Automated Vehicles Based on Extreme Value Theory

This paper proposes a proactive fleet monitoring approach for automated vehicles (AVs) based on Extreme Value Theory (EVT) to reduce the accident risk during first deployment and software updates. By performing sequential statistical tests on threat metrics measured in an AV fleet, the monitor can be used to quickly identify and abort operations if the AVs do not meet the required level of safety. To evaluate the proposed monitoring approach, it is studied in a fictive deployment case using two different threat metrics, one predictive and one retrospective. The evaluation showed that a significant risk reduction is achievable when using the EVT fleet monitor compared to reactive fleet monitoring. Moreover, using a predictive threat metric reduces the risk of accidents dramatically. However, it has the drawback of frequently aborting operations unless the systems are significantly better than required. On the other hand, a retrospective threat metric was a more balanced alternative that could substantially reduce risk without being overly conservative. In summation, the EVT fleet monitoring is a promising complementary approach to traditional validation to minimize the risk and abort operations of sub-performing new deployments well before any accidents are caused.